AI traffic in GA4: 1% of sessions, and GA4 files it wrong

"How many visitors are we getting from ChatGPT?" It has become the first question in every marketing meeting, and the answer rarely survives two minutes before turning into an argument about method.

So we stopped guessing. We queried the Google Analytics Data API across 43 GA4 properties spread over 8 accounts, 40 of them usable, for 2,085,299 sessions over twelve complete months, from 1 August 2025 to 31 July 2026.

The result fits in three lines, and two of them contradict what everyone publishes. Traffic from answer engines is 1.02% of total traffic. It has not moved all year. And GA4 files it so badly that four counting methods out of five lose part of it on the way.

What was measured, and what was not

The panel is described without makeup, because it governs how every number below should be read.

• 43 GA4 properties across 8 accounts. 40 return at least one session, 3 are empty, none failed.
• Unbalanced composition: 20 of the 40 usable properties are hospitality brochure sites belonging to a single account, 19 are B2B sites spread over 6 accounts, the last one is vydera.com. Any unweighted average would tilt towards the brochure sites, so the aggregate figures below are weighted by sessions.
• Every property is anonymised except vydera.com. This is client data: it can be measured, it cannot be named.
• Method: Analytics Data API v1beta, runReport, dimensions sessionSource, sessionMedium, yearMonth and sessionDefaultChannelGroup, plus a pageReferrer probe. Every source value was pulled down and classified locally, with no guessed server-side filter.

And here is what the reading does not say, so what this article will never claim. Traffic from Google AI Overviews is not measurable: it is folded into google / organic and no GA4 dimension separates it. Invisible, not absent. Clicks from the desktop and mobile apps of ChatGPT, Perplexity or Claude send neither referrer nor utm: they land in (direct) and cannot be attributed to anything. No estimate of their volume was attempted. Finally, the economic value of this traffic was not measured: GA4 key events are not sales. August 2026, still in progress, is deliberately excluded from the window since GA4 does not consolidate the last 48 hours, and the split by country or device was not queried.

1.02% of traffic, and two sectors that look alike

Over twelve months, 21,168 sessions come from an answer engine, out of 2,085,299. That is 10.15 per thousand, or 1.02%.

The spread is wide: from 1.37 to 26.68 per thousand depending on the property, counting only those above 500 sessions in the window. A factor of 19 between the least and the most exposed site in the panel.

The original brief promised a sector breakdown. Once anonymised, it honestly comes down to two buckets, and they say almost nothing:

• 19 B2B sites, 1,703,488 sessions, 17,382 AI sessions, or 10.20 per thousand.
• 20 hospitality brochure sites, 376,986 sessions, 3,726 AI sessions, or 9.88 per thousand.

Three tenths of a point per thousand apart. Sector explains nothing here: the spread inside each bucket dwarfs the gap between them. An article selling you an AI traffic share "by sector" is selling you the average of two overlapping clouds.

Three properties recorded no AI session at all. They have 1, 57 and 223 sessions over twelve months: that is an empty sample, not a result. Nothing can be concluded from it, least of all that three sites in forty get nothing from AI engines.

As for vydera.com, the only nameable property: 4,825 sessions, 60 AI sessions, 12.44 per thousand, over 7 months only, since the property was created on 28 January 2026. First AI session in March 2026. That sits above the panel median, and on 60 sessions it licenses no trend whatsoever.

AI traffic is not growing

This is the most uncomfortable result, and it is unambiguous.

To stop a property joining or leaving mid-window from manufacturing a slope, the trend is computed on a fixed cohort of 31 properties, the ones that returned all twelve months. On that cohort, the AI share moves from 10.02 per thousand in the first half-year to 10.24 in the second.

Nought point two two per thousand of difference, when monthly noise runs from 8.63 to 11.48. The high point is November 2025, the low point December 2025: that is monthly noise, not a trajectory.

On this panel, over this window, the share of traffic coming from AI engines is flat. It says nothing about the absolute volume of questions asked to assistants, which is not measured here. It says the fraction of visits they send to a site has not changed in twelve months.

And it does not behave better than organic

The other standard promise is that a visitor arriving from an AI engine is better qualified. We compared, property by property and never pooled, AI sessions against the Organic Search channel of the same property. Threshold: at least 30 AI sessions. 33 properties clear it.

• Average duration: median AI-to-organic ratio 0.903. 12 properties out of 33 show a longer AI duration.
• Pages per session: median 1.027. 18 out of 33 above.
• Engagement rate: median 1.014. 18 out of 33 above.

Eighteen out of thirty-three is a coin toss. No consistent direction emerges. And this has to be said plainly: these are counts, no statistical test was run, and 12 out of 33 or 18 out of 33 are not significance levels.

On conversion, 28 properties have key events configured. The median AI-to-organic ratio there is 0.668, with 11 properties out of 28 above 1. But the range runs from 0 to 7.63, and a GA4 key event is not a sale: its definition changes from one property to the next. That ratio should never be quoted without its range and without that caveat. If you are trying to settle on metrics that actually decide something, we wrote a separate piece on it.

ChatGPT is two thirds of it, Copilot is everywhere and weighs nothing

The breakdown by engine, in sessions and in number of properties reached:

• ChatGPT: 14,101 sessions across 37 properties
• Perplexity: 4,249 sessions across 31 properties
• Gemini: 1,548 sessions across 28 properties
• Claude: 832 sessions across 19 properties
• Copilot: 331 sessions across 26 properties
• Le Chat 55, Kagi 33, Poe 13, Grok 12, DeepSeek 11, Felo 2

Two readings come out of that list. First: ChatGPT accounts for two thirds of the panel's AI traffic, and it reaches 37 properties out of 40. The second one is subtler. Copilot shows up on 26 properties for 331 sessions in total, meaning near-universal presence and negligible volume. Presence and volume are not summarised by the same number, and a dashboard showing only how many engines were detected tells you nothing useful.

Eight engines from the starting list produced no session at all across the 43 properties: Andi, Exa, Genspark, Komo, Meta AI, Phind, You.com and iAsk. And the long tail of niche interfaces actually observed, nine sources from Qwen to Duck.ai, is worth 38 sessions between them. Watching those engines costs time and returns nothing.

Where GA4 actually files AI traffic

Here is the heart of the problem, and the reason two people measuring the same site come back with two different numbers. Over twelve months, the 21,168 AI sessions land in GA4 channels like this:

• Referral: 12,705 sessions, roughly 60%
• Unassigned: 5,714 sessions, roughly 27%
• AI Assistant: 2,724 sessions, roughly 13%
• Organic Search: 44 sessions

More than a quarter of AI traffic sits in Unassigned, the channel nobody ever opens. And the native "AI Assistant" channel, the one Google finally shipped and everyone was waiting for, covers 12.9% of the window.

The reason is simple and nobody states it: that channel does not exist before June 2026, and it is not retroactive. Across the first ten months of the reading it shows zero sessions. In June 2026 it covers 74.6% of the referrer reading, in July 97.8%. Across twelve months, 12.9%. An annual report built on that channel undercounts AI traffic by a factor of 7.8.

The June 2026 trap: the medium switched without warning

Same month, different effect, and this one breaks dashboards that already exist. The medium label carried by AI sessions changed:

• May 2026: referral 1,129, (not set) 750, ai-assistant 0
• June 2026: ai-assistant 1,410, referral 297, (not set) 173
• July 2026: ai-assistant 1,314, referral 16, (not set) 12

In other words, every segment built on "medium = referral" stopped working in June 2026. No warning, no message, nothing. The chart does not show an error: it shows a drop. Plenty of people probably read it as AI traffic falling, when it was a relabelling.

The problem is older than that, incidentally. Over twelve months, chatgpt.com splits into referral 7,095, (not set) 4,840, ai-assistant 2,110, organic 41 and (none) 7. A segment on medium referral therefore catches only half of ChatGPT sessions, and that was already true before the switch: sessions arriving with a utm_source but no utm_medium fall into (not set).

Same trap on the source side. The bare source perplexity, with no dot and no domain, produces 832 sessions across 28 properties, 825 of them on medium (not set). A segment targeting only perplexity.ai therefore leaves 20% of Perplexity traffic outside. The rule that follows is simple: never segment on the medium, and never assume the spelling of a source.

The pageReferrer filter misses up to 43% and double counts

The recipe most blogs recommend is to filter on pageReferrer containing chatgpt.com. We probed it on the four properties receiving the most AI traffic in the panel, comparing it to the sessionSource reading for the same engine:

• 3,328 sessions against 4,878, so 32% missed
• 832 against 1,185, so 30% missed
• 487 against 851, so 43% missed
• 459 against 745, so 38% missed

The cause is mechanical: pageReferrer is empty as soon as the referrer was not passed along, which happens often in private browsing, in some apps and behind certain redirects.

But the worst part is elsewhere. The same filter adds 528, 150, 87 and 95 internal echo sessions on those four properties. The mechanism is worth understanding: a visitor arrives from ChatGPT on a URL carrying ?utm_source=chatgpt.com, clicks an internal link, and that page of the site becomes the referrer of the next one. The filter counts it as a fresh arrival from ChatGPT. That is pure double counting.

Net result: a method that misses a third of the real traffic and partly replaces it with manufactured traffic. It can land on the right number by cancellation, which is the worst possible outcome.

The segment that works, and the list of 26 sources

The only method that captures everything in this reading relies on the sessionSource dimension, with a regular expression built on sources actually observed, not on a copied list.

The list below separates two circles. The canonical circle gathers consumer assistants under every spelling, including bare utm_source values such as perplexity and openai. The long tail gathers nine niche interfaces worth 38 sessions in total. Referrers ending in .ai that are business applications rather than answer engines were deliberately cut: without that cut, the list fills up with SaaS tools and the number inflates for nothing.

Running the measurement yourself

All of this replays with the Analytics Data API v1beta and a read-only OAuth token. Four method points, every one of them learned the hard way.

1. Pull every sessionSource value down, then classify locally. A server-side filter on a guessed list will make you miss the spellings you did not anticipate, and you will never know which ones.
2. Take numerator and denominator from the same split. On this panel, the undivided total returns 2,103,605 sessions while the sum of the twelve months returns 2,085,299: a 0.88% gap. That is GA4 approximating when sessions are split by a dimension, not a scripting error, and it is enough to shift a share by a few percent.
3. Drop the retention fear. It is the most widespread belief and it is wrong: 26 properties in the panel are set to 2 months of event data retention, and 18 of them still returned all twelve months. The retention setting blocked no Data API measurement on standard dimensions. Not one empty month in this reading is explained by retention.
4. Cross-check before publishing. Three controls were run here: no (other) row appeared across the 43 properties, so no cardinality collapse distorts the denominators; the sum of the 31 days of July 2026 on the largest AI property returns 337 sessions against 337 for the monthly total, zero gap; and an independent cross-check on vydera.com, through a different path, returns 60 sessions against 60 with an identical breakdown down to the character.

Last point, and it conditions all the others: the reading covered all 43 properties across the 8 accounts with zero properties in error, after retrying the network drops. A portfolio that returns a partial result is not a measured portfolio. Until the error counter hits zero, the denominators are wrong and nothing tells you so.

What to do with it

AI traffic is not, today and on this panel, a volume channel. At 1% of traffic and with no growth over twelve months, it will fund no media plan. Running it as an acquisition channel means committing to commentary on monthly noise.

What it is instead is a proof channel: every AI session attests that an answer engine cited your page inside a conversation. That is why the metric that counts is not the session but the citation rate, and why both numbers belong together in an SEO and AEO dashboard, never apart.

Three moves, in order. One: replace every medium-based segment with the sessionSource segment above, today, before you discover the June 2026 switch inside a quarterly report. Two: add the four channels together, Referral, Unassigned, AI Assistant and Organic Search, or you will leave a quarter of the traffic in a drawer. Three: date your series. A channel that does not exist before June 2026 cannot be compared to May.

And if you want someone to wire up the measurement and then move the number, that is exactly what our performance tracking offer does.